Sumo Action Classification Using Mawashi Keypoints

Yuki Utsuro

, Hidehiko Shishido

and Yoshinari Kameda

Master's Program in Intelligent and Mechanical Interaction Systems,

University of Tsukuba, Tennodai 1-1-1, Tsukuba, Ibaraki, Japan

Center for Computational Sciences, University of Tsukuba, Tennodai 1-1-1, Tsukuba, Ibaraki, Japan

Keywords: Action Classification, Action Recognition, Extended Skeleton Model, Pose Estimation, Pose Sequences,

Computer Vision in Sports, Sumo, Japanese Wrestling.

Abstract: We propose a new method of classification for kimarites in sumo videos based on kinematic pose estimation.

Japanese wrestling sumo is a combat sport. Sumo is played by two wrestlers wearing a mawashi, a loincloth

fastened around the waist. In a sumo match, two wrestlers grapple with each other. Sumo wrestlers perform

actions by grabbing their opponents' mawashi. A kimarite is a sumo winning action that decides the outcome

of a sumo match. All the kimarites are defined based on their motions. In an official sumo match, the kimarite

of the match is classified by the referee, who oversees the classification just after the match. Classifying

kimarites from videos by computer vision is a challenging task. There are two reasons. The first reason is that

the definition of kimarites requires us to examine the relationship between the mawashi and the pose. The

second reason is the heavy occlusion caused by the close contact between wrestlers. For the precise

examination of pose estimation, we introduce a wrestler-specific skeleton model with mawashi keypoints.

The relationship between mawashi and body parts is uniformly represented in the pose sequence with this

extended skeleton model. As for heavy occlusion, we represent sumo actions as pose sequences to classify

the sumo actions. Our method achieves an accuracy of 0.77 in action classification by LSTM. We confirmed

that the skeleton model extension by mawashi keypoints improves the accuracy of action classification in

sumo through the experiment results.

1 INTRODUCTION

Japanese wrestling sumo is a two-person combat

sport. Sumo wrestlers wear a mawashi, a loincloth

fastened around their body. Two wrestlers grapple

and fight against each other in a circular sumo ring

called dohyo. They are allowed to hold the

opponent’s mawashi to control the opponent's body.

Therefore, during a sumo match, the wrestlers are in

close contact with each other (Figure 1). This contact

causes heavy occlusion. One of the unique features of

sumo is the mawashi. The mawashi is tightened so

that it adheres closely to the wrestler's body.

A kimarite is a sumo winning action that decides

the outcome of a sumo match. All the kimarites are

defined based on their motions. The definition of

sumo actions involves the positional relationship

https://orcid.org/0009-0008-7166-2235

https://orcid.org/0000-0001-8575-0617 (Currently at

Soka University)

https://orcid.org/0000-0001-6776-1267

between the hands of one wrestler and the mawashi

of the other. Therefore, examination of this

relationship is an indispensable property for

discriminating actions. In an official sumo match, the

kimarite of the match is classified by the referee, who

oversees the classification just after the match.

Classifying kimarites from videos by computer

vision is a challenging task. There are two reasons for

this. The first reason is that the definition of kimarites

requires us to examine the relationship between the

mawashi and the pose. The second reason is the heavy

occlusion caused by the close contact between

wrestlers.

We propose a new method of classification for

kimarites in sumo videos based on kinematic pose

estimation. For the precise examination of pose

estimation,

newly

introduce a

wrestler-specific

Utsuro, Y., Shishido, H. and Kameda, Y.

Sumo Action Classiﬁcation Using Mawashi Keypoints.

DOI: 10.5220/0012337000003660

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 19th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2024) - Volume 2: VISAPP, pages

401-408

ISBN: 978-989-758-679-8; ISSN: 2184-4321

401

skeleton model with mawashi keypoints. The

relationship between mawashi and body parts is

uniformly represented as the pose of this extended

skeleton model. As for handling heavy occlusion, we

represent sumo actions as pose sequences to classify

the sumo actions so that the later classification

procedure can achieve better classification

performance.

The first contribution of our research is the new

skeleton model with the mawashi keypoints. Since

the mawashi is tightened to the wrestler's body, it can

be regarded as a part of the body. We extended the

existing human skeleton model in OpenPose (Cao et

al., 2019) by adding two new keypoints on the

mawashi worn by the wrestler (Figure 2). This

extension makes it possible to uniformly represent the

relationship between the mawashis and the body parts

within a single pose.

The second contribution is the utilization of pose

sequences for the representation of sumo actions so

that the kimarite classification procedure can achieve

better classification performance. The kimarite, the

last action of a match that decides the outcome of the

match, lasts approximately one-second movement of

the wrestlers. Sumo referees can classify the action by

observing the one-second movement. Even in

situations with heavy occlusions during matches,

action classification should be successfully made by

referring to the pose sequence that might miss some

parts in some poses. The pose sequences are classified

by LSTM (Hochreiter and Schmidhuber, 1997).

Figure 1: An overview of sumo. The loincloths that wrestlers

wear are called "mawashi." The ring is called "dohyo."

2 RELATED WORKS

2.1 Action Classification in Sports

Methods for action classification from specific sports

videos often differ from those for daily-action

classification due to the need to consider the unique

characteristics of the sport. Regarding action

classification in sports videos, studies have been

reported focusing either on a single player or multiple

players. Studies that focus on the action of just one

player in sports are reported across various sports

such as table tennis (Kulkarni and Sheony, 2021),

basketball (Zakharchenko, 2020; Liu et al., 2021), ice

hockey (Fani et al., 2017; Vats et al., 2019), karate

(Guo et al., 2021), taekwondo (Liang and Zuo, 2022;

Luo et al., 2023), figure skating (Liu et al., 2021; Liu

et al., 2020), and tennis (Vainstein et al., 2014). In

these sports, little occlusion occurs between players.

These studies show that if the action is made by a

single player and little occlusion is found in the sports

video, it is feasible to classify actions.

There are also studies on multi-player action

classification in some sports, such as volleyball

(Gavrilyuk et al., 2020; Azar et al., 2019), hockey

(Rangasamy et al., 2020), soccer (Gerats, 2020), and

badminton (Ibh et al., 2023). These studies use

information from multiple players as input, but no

heavy occlusions occurred for the players under

observation. This is because close contact between

players does not frequently happen in these sports.

In sports where two athletes are in close contact,

leading to heavy occlusion, a study has been reported

on action classification in freestyle wrestling

(Mottaghi et al., 2020). This study classifies actions

using features extracted from the silhouette of

grappling wrestlers. This method employs SVM or k-

NN as a classifier, which might not consider the

temporal changes adequately. Since it does not

perform pose estimation based on the human skeleton,

this approach may not be suitable for tasks that

require more detailed observation of body

movements.

Regarding sumo research using machine learning,

there is a study (Li and Sezan, 2001) that estimates

the start timing of the match using SVM. However,

this is not a study of action classification.

2.2 Extension of Human Skeleton

Model for Pose Estimation

In the field of pose estimation in sports, there have

been reports of research using skeleton models that

have been extended by adding sport-specific

keypoints to the existing human skeleton model.

Neher et al. have proposed an extended skeleton

model that has two new keypoints of the upper and

lower parts of an ice hockey stick (Neher et al., 2018).

They show that the introduction of this extended

skeleton model improves the accuracy of pose

estimation for ice hockey players.

Ludwig et al. have proposed a pose estimation

method using an extended skeleton model that

incorporates

two endpoints of the ski board in ski

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

402

Figure 2: Proposed extended skeleton model. (Left) Original skeleton model of OpenPose (Cao et al., 2019). This picture is

cited from the OpenPose document webpage (CMU Perceptual Computing Lab, 2021). (Middle) Our extended skeleton model.

Keypoint no. 25 and no. 26 are the new keypoints on mawashi. (Right) Pose estimation with the extended skeleton model.

Keypoints highlighted by yellow circles are mawashi keypoints.

jumping as the keypoints (Ludwig et al., 2020). They

have reported that this model improves the accuracy

of estimating flight parameters in ski jumping videos.

Einfalt et al. have proposed an extended skeleton

model based on a skeleton model without foot

keypoints. They added four points of the left and right

toes and the heels for track and field athletes (Einfalt

et al., 2019). They have confirmed that this model

improves the accuracy of frame-level estimation of the

timing of steps in long jump and triple jump videos.

These studies show that adding unique keypoints

to the existing skeleton model improves the accuracy

of pose-based action recognition tasks in sports.

3 SUMO ACTIONS “KIMARITE”

The Japan Sumo Association officially recognizes 82

kimarites. In this study, we focus on classifying the

most popular four kimarites based on the frequency

of kimarite appearance over the past five years (Japan

Sumo Association, 2023). There are Oshidashi

(Frontal Push Out), Yorikiri (Frontal Force Out),

Hatakikomi (Slap Down), and Tsukiotoshi (Thrust

Down). These top four kimarite actions account for

approximately 65% of the total matches. These

kimarites are shown in Figure 3. Usually, kimarites

last one second or so.

4 SUMO ACTION

CLASSIFICATION

4.1 Procedure Overview

We assume sumo videos used for classification are

taken from a single viewpoint, where the camera

captures both the ring and the wrestlers. The video is

taken from a perspective where the two wrestlers are

positioned on the left and right in the frame at the start

of the match. Additionally, the video is taken from a

sufficiently distant place from the ring, such that it

can be approximated by a weak perspective

projection (as shown in Figure 4). Note that the

camera changes the pan/tilt/zoom parameters slightly

to let the wrestlers in the video.

Figure 3: Four kimarite actions we classify. Upper-left is

Oshidashi (Frontal Push Out), upper-right is Yorikiri

(Frontal Force Out), lower-left is Hatakikomi (Slap Down),

and lower-right is Tsukiotoshi (Thrust Down). These

pictures are cited from the official website with the courtesy

of the Japan Sumo Association (Japan Sumo Association,

2023).

Figure 4: A snapshot of sumo videos used for classification.

All videos are captured like this figure.

When given an input video, we first perform the

frame-by-frame pose estimation and detect the

keypoints of the sumo wrestlers' skeleton. Our

approach is based on OpenPose (Cao et al., 2017; Cao

et al., 2019). We employ the extended skeleton model

(Figure 2) tuned for sumo wrestlers, which consists of

Sumo Action Classiﬁcation Using Mawashi Keypoints

403

27 keypoints, including two new keypoints on

mawashi. Using this model for pose estimation, we

obtain the sets of skeleton positional vectors for the

two sumo wrestlers.

These two skeleton positional vectors are

concatenated into one integrated feature vector as a

pose sequence.

Since sumo takes place within the ring, the

coordinates should be normalized to the ring. Since

the ring is a circle shape, it is expressed by the planner

cartesian coordinate system with the range of -1.0 to

1.0 for both axes.

We train the LSTM network with a normalized

pose sequence (Hochreiter and Schmidhuber, 1997)

by using the class name of the kimarite as the correct

label.

4.2 Extended Skeleton Model

We introduce a new extended 27-point skeleton

model (Figure 2-middle) based on the existing 25-

point model of OpenPose (Figure 2-left) by adding

two new keypoints of the mawashi. The added

keypoints represent the knot point between the front

mawashi and the front vertical flap (Figure 5-A) and

the knot point between the back mawashi and the

back vertical flap (Figure 5-B).

We created a custom dataset of 723 images from

sumo matches with keypoint annotations of the sumo

wrestlers. Using this dataset, we fine-tuned our

extended skeleton model with the pretrained

parameters of the 25-point model as the initiation.

By using the extended skeleton model, it becomes

possible to represent the relationship between the

mawashi and body parts such as hands, which is

essential for classifying sumo actions in the

uniformed form of skeleton keypoints.

Figure 5: Two additional keypoints of mawashi.

4.3 Normalization to the Sumo Ring

We obtain the planner coordinates of the skeleton

keypoints of the two wrestlers by applying the

extended skeleton model to a video frame. Since the

camera moves slightly, the coordinates should be

normalized to the circular sumo ring. Since the ring is

observed as an ellipse in the frame, an ellipse

approximation is used in image processing to extract

the circle information of the sumo ring. We normalize

the coordinates of the keypoints to the planner

cartesian coordinate system with the range of -1.0 to

1.0 for both axes. The two axes are set by the

definition of the sumo ring as it is placed to align

north, south, east, and west.

4.4 Pose Sequence

In sumo, the class of the kimarite is determined based

on the movements made by the two wrestlers from the

start of the match-finishing action to its end. When

classifying sumo actions, it is essential to consider the

movements of both wrestlers. We represent the

kimarite action as time series data related to the

movements of the two wrestlers. We first concatenate

the two sets of the planner coordinates of the skeleton

keypoints of the two wrestlers into one frame vector.

Then, by arranging the frame vectors in time-series

order, we obtain a pose sequence that corresponds to

a certain time length of the sumo match.

4.5 Training LSTM

As vision-based action classifications are one of the

popular research domains, research such as graph

convolution (Shi et al., 2019), Vision Transformer

(Dosovitskiy et al., 2020), ViViT (Arnab et al., 2021),

and TemPose (Ibh et al., 2023) have been reported.

As the pose sequence of kimarite is not long, we

would rather adopt LSTM, one of the RNN-based

methods.

We make a classifier model composed of an

LSTM layer, Dropout layers, and Dense layers. The

structure of the model is shown in Figure 6. The

model has two Dense layers and three Dropout layers

between the first LSTM layer and the last Dense

layer. These layers are set to prevent overfitting and

stabilize the learning process.

Figure 6: The classifier model.

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

404

5 EXPERIMENT

5.1 Dataset

We take up four classes of kimarite action; Oshidashi

(Frontal Push Out), Yorikiri (Frontal Force Out),

Hatakikomi (Slap Down), and Tsukiotoshi (Thrust

Down). We made a video dataset collected from the

NHK Sports webpage (NHK, 2023) for our

experiments. All videos were taken by a camera

placed at a certain distance from the sumo ring. The

resolution of the videos is 960 pixels in width and 540

pixels in height at 30 fps.

Kimarite is the last action of the match. Therefore,

as an input video clip for classification, we segment

the last 20 frames of the match video. The end frame

of the match is marked visually. The numbers of the

collected video clips for the four kimarites are shown

in Table 1. The constructed dataset is randomly split

so that the ratio of training data to test data is 7:3.

Table 1: The numbers of kimarites in the dataset.

Action # of video cli

Hatakikomi (Slap Down) 223

Oshidashi (Frontal Push Out) 235

Tsukiotoshi

(

Thrust Down

)

209

Yorikiri

(

Frontal Force Out

)

251

5.2 Data Augmentation

Since the number of video clips is not large, we apply

data augmentation to the training video clips. The

data augmentation should be tailored to fit the sumo

property.

5.2.1 Random Change

As for the general data augmentation, we randomly

add a small positional noise to the original training

dataset. The noise is generated based on the standard

normal distribution. An example of the random

change is shown in Figure 7. The red and grey

skeletons are the original, and the pink and yellow

skeletons are the slightly changed ones in the figure.

5.2.2 Horizontal Flip

For all sumo matches, the name of the kimarite action

remains the same even if the two sumo wrestlers are

horizontally flipped around the centre of the sumo

ring. In Figure 8, the pink and yellow skeletons on the

right are the flipped skeletons of the red and grey

original skeletons on the left. We doubled the number

of pose sequence data for training by the horizontal

flip.

5.2.3 Relocation

The kimarite actions of Hatakikomi (Slap Down) and

Tsukiotoshi (Thrust Down) should end within the

sumo ring. Therefore, these two actions are still valid

even when the actions are relocated within the sumo

ring.

We generated new pose sequences with relocation

by adding the same values to all the coordinates in a

pose sequence. The random values are generated

based on the uniform distribution. The values are set

so as not to go over the sumo ring.

This process allows us to create multiple new

skeletons while maintaining the shape of the original

skeleton, as shown in Figure 9. In the figure, the red

and grey skeletons represent the original skeletons,

while the pink and yellow skeletons, as well as the

light blue and lavender skeletons, are the relocations

of the original skeletons.

Figure 7: Random change of skeleton poses. The blue circle

indicates the sumo ring.

Figure 8: Horizontal flipping of skeleton poses. The blue

circle indicates the sumo ring.

Figure 9: Relocation of skeleton poses. The blue circle

means the sumo ring.

Sumo Action Classiﬁcation Using Mawashi Keypoints

405

5.2.4 Enhanced Training Dataset

On kimarite classification, we do not need to

distinguish the winner and the loser of a match.

Therefore, we created pose sequences by swapping

the two sumo wrestlers in the pose sequences of all

the video clips.

We applied all the augmentation approaches and

prepared the enhanced training dataset. The final

numbers of the pose sequences in the enhanced

training dataset are shown in Table 2.

Table 2: Pose sequences of the enhanced training dataset.

Action # of

ose se

Hatakikomi (Slap Down) 3099

Oshidashi (Frontal Push Out) 3263

Tsukiotoshi

(

Thrust Down

)

2903

Yorikiri

(

Frontal Force Out

)

2632

5.3 Model Training

We trained the model shown in Figure 6. The loss

function for the model is cross-entropy. Layer

normalization is performed after the LSTM layer. All

the activation functions in the Dense layer are ReLU,

and the dropout rate in the Dropout layer is set at 0.2.

30% of the video clips were randomly selected for

validation. Training was carried out for 50 epochs,

and the weights from the epoch with the lowest

validation loss were adopted.

The training loss and training accuracy are shown

in Figure 10, and the validation loss and validation

accuracy are shown in Figure 11.

Figure 10: Training loss and training accuracy.

Figure 11: Validation loss and validation accuracy.

5.4 Results and Evaluations

5.4.1 Evaluation of Our Method

As the evaluation of the trained model, we performed

classification with the test dataset. As mentioned in

section 5.1, this test dataset is randomly picked 30%

of the overall dataset shown in Table 1. We calculate

the precision, recall, and F1-score of the results of

classification. The results are shown in Table 3.

Additionally, the confusion matrix is shown in Figure

12. Figure 12 shows that the correctly predicted class

is the most common for any of the classes. The

accuracy for classifying the four kimarite actions was

0.77. Based on these results, our method is effective

in classifying kimarite actions from sumo videos.

From Figure 12, we can confirm that the actions

tend to be misclassified into certain specific classes.

Specifically, Hatakikomi are mostly misclassified

as Tsukiotoshi, and vice versa. A similar case can be

found for the pair Oshidashi and Yorikiri. The pair of

Hatakikomi and Tsukiotoshi look similar from the

definition (lower row of Figure 3). The pair of

Oshidashi and Yorikiri (upper row of Figure 3) also

look similar. This implies that most of the

misclassifications occur between actions that have

similar forms. We expect that further training of the

LSTM model with a larger number of the dataset

could result in better recognition performance.

5.4.2 Improvement by Mawashi Keypoints

To verify the effectiveness of adding the mawashi

keypoints, we compared the results of the 25-point

model (Figure 2), which did not include the two

extended mawashi keypoints. The result of the 25-

point model is presented in Table 4, and the confusion

matrix is shown in Figure 13. Furthermore, the result

of the comparison of accuracy is shown in Table 5.

As shown in Table 5, using the extended skeleton

model with the mawashi keypoints is more effective

for kimarite action classification. From this result, we

can confirm the advantage of the extended skeleton

model tuned for sumo wrestlers.

Table 3: Results of kimarite action classification using the

proposed extended skeleton model with mawashi

keypoints. The total accuracy of the 4-class classification is

0.77.

Action Precision Recall F1-score

Hatakikomi 0.72 0.84 0.77

Oshidashi 0.78 0.76 0.77

Tsukiotoshi 0.78 0.63 0.70

Yorikiri 0.78 0.82 0.80

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

406

Table 4: Results of kimarite action classification using the

25-point skeleton model without mawashi keypoints. Total

accuracy is 0.74, as shown in Table 5.

Action Precision Recall F1-score

Hatakikomi 0.70 0.79 0.74

Oshidashi 0.86 0.69 0.77

Tsukiotoshi 0.65 0.62 0.63

Yorikiri 0.77 0.86 0.81

Figure 12: Confusion matrix of the action classification

result using the proposed extended skeleton model with

mawashi keypoints.

Figure 13: Confusion matrix of the action classification

result using the 25-point skeleton model without mawashi

keypoints.

Table 5: Comparison of classification accuracy between

two skeleton models (with mawashi keypoints or not).

Skeleton model Accurac

Proposed: With mawashi keypoints

(

oints

)

0.77

Without mawashi keypoints

(25 points)

0.74

6 CONCLUSIONS

We propose a new method of classification for

kimarites in sumo video based on our proposed

extended skeleton model that has the two mawashi

keypoints to fit sumo wrestlers.

The unique feature of the extended skeleton

model is the new keypoints that correspond to

mawashi. The relationship between mawashi and

body parts is uniformly represented in the extended

skeleton model. As to deal with heavy occlusion, we

represent sumo actions as pose sequences so that the

classification procedure based on LSTM can achieve

better classification performance.

The extended skeleton model is trained by our

own video dataset, which was enhanced with the data

augmentation. The data augmentation was designed

so as not to break the definition of kimarites.

With this approach, we achieved an accuracy of

0.77 in the task of classifying the four popular

kimarites. We also confirmed that the introduction of

the extended skeleton model with mawashi keypoints

improved classification accuracy by the ablation

study with the 25-point model.

ACKNOWLEDGEMENTS

We thank the Japan Sumo Association for the

permission to use their images in this paper. We had

a written permission from them. Part of this work was

supported by JSPS KAKENHI Grant Number

22K19803.

REFERENCES

Arnab, A., Dehghani, M., Heigold, G., Sun, C., Lučić, M.,

and Schmid, C. (2021). Vivit: A video vision

transformer. In Proceedings of the IEEE/CVF

international conference on computer vision, pp.6836-

6846.

Azar, S. M., Atigh, M. G., Nickabadi, A., and Alahi, A.

(2019). Convolutional relational machine for group

activity recognition. In Proceedings of the IEEE/CVF

Conference on Computer Vision and Pattern

Recognition, pp.7892-7901.

Cao, Z., Hidalgo, G., Simon, T., Wei, S. E., and Sheikh, Y.

(2019). OpenPose: Realtime Multi-Person 2D Pose

Estimation Using Part Affinity Fields. In IEEE

Transactions on Pattern Analysis & Machine

Intelligence, vol.43, no.1, pp.172-186.

Cao, Z., Simon, T., Wei, S. E., and Sheikh, Y. (2017).

Realtime multi-person 2d pose estimation using part

affinity fields. In Proceedings of the IEEE conference

on computer vision and pattern recognition (CVPR),

pp.7291-7299.

CMU Perceptual Computing Lab. (2021). OpenPose Doc –

Output, OpenPose Doc, Available at:

https://github.com/CMU-Perceptual-Computing-

Lab/openpose/blob/master/doc/02_output.md

(Accessed: 10 October 2023).

Sumo Action Classiﬁcation Using Mawashi Keypoints

407

Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn,

D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer,

M., Heigold, G., Gelly, S., Uszkoreit, J., and Houlsby,

N. (2020). An image is worth 16x16 words:

Transformers for image recognition at scale. arXiv

preprint arXiv:2010.11929.

Einfalt, M., Dampeyrou, C., Zecha, D., and Lienhart, R.

(2019). Frame-level event detection in athletics videos

with pose-based convolutional sequence networks. In

Proceedings Proceedings of the 2nd International

Workshop on Multimedia Content Analysis in Sports,

pp.42-50.

Fani, M., Neher, H., Clausi, D. A., Wong, A., and Zelek, J.

(2017). Hockey action recognition via integrated

stacked hourglass network. In Proceedings of the IEEE

conference on computer vision and pattern recognition

workshops, pp.29-37.

Gavrilyuk, K., Sanford, R., Javan, M., and Snoek, C. G.

(2020). Actor-transformers for group activity

recognition. In Proceedings of the IEEE/CVF

Conference on Computer Vision and Pattern

Recognition, pp.839-848.

Gerats, B. G. A. (2020). Individual action and group

activity recognition in soccer videos. Master's thesis,

University of Twente.

Guo, J., Liu, H., Li, X., Xu, D., and Zhang, Y. (2021). An

attention enhanced spatial–temporal graph

convolutional LSTM network for action recognition in

karate. Applied Sciences, vol.11, no.18, article no.8641.

Hochreiter, S., and Schmidhuber, J. (1997). Long short-

term memory. Neural computation, vol.9, no.8,

pp.1735-1780.

Ibh, M., Grasshof, S., Witzner, D., and Madeleine, P.

(2023). TemPose: A New Skeleton-Based Transformer

Model Designed for Fine-Grained Motion Recognition

in Badminton. In Proceedings of the IEEE/CVF

Conference on Computer Vision and Pattern

Recognition, pp.5198-5207.

Japan Sumo Association. (2023). 82 kimarite. Nihon Sumo

Kyokai Official Grand Sumo Home Page, Available at:

https://www.sumo.or.jp/Kimarite (Accessed: 10

October 2023).

Kulkarni, K. M., and Shenoy, S. (2021). Table tennis stroke

recognition using two-dimensional human pose

estimation. In Proceedings of the IEEE/CVF

conference on computer vision and pattern recognition,

pp.4576-4584.

Li, B., and Sezan, M. I. (2001). Event detection and

summarization in sports video. In Proceedings IEEE

workshop on content-based access of image and video

libraries (CBAIVL 2001), pp.132-138. IEEE.

Liang, J., and Zuo, G. (2022). Taekwondo Action

Recognition Method Based on Partial Perception

Structure Graph Convolution Framework. Scientific

Programming, vol.2022, Article ID 1838468.

Liu, R., Liu, Z., and Liu, S. (2021). Recognition of

basketball player’s shooting action based on the

convolutional neural network. Scientific Programming,

vol.2021, Article ID 3045418.

Liu, S., Liu, X., Huang, G., Feng, L., Hu, L., Jiang, D.,

Zhang, A., Liu, Y., and Qiao, H. (2020). FSD-10: a

dataset for competitive sports content analysis. arXiv

preprint arXiv:2002.03312.

Liu, S., Zhang, A., Li, Y., Zhou, J., Xu, L., Dong, Z., and

Zhang, R. (2021). Temporal segmentation of fine-

gained semantic action: A motion-centered figure

skating dataset. In Proceedings of the AAAI conference

on artificial intelligence, vol.35, no.3, pp.2163-2171.

Ludwig, K., Einfalt, M., and Lienhart, R. (2020). Robust

estimation of flight parameters for ski jumpers. In 2020

IEEE International Conference on Multimedia & Expo

Workshops (ICMEW), pp.1-6. IEEE.

Luo, C., Kim, S. W., Park, H. Y., Lim, K., and Jung, H.

(2023). Agnostic Taekwondo Action Recognition

Using Synthesized Two-Dimensional Skeletal Datasets.

Sensors, vol.23, no.19, p.8049.

Mottaghi, A., Soryani, M., and Seifi, H. (2020). Action

recognition in freestyle wrestling using silhouette-

skeleton features. Engineering Science and Technology,

an International Journal, vol.23, no.4, pp.921-930.

Neher, H., Vats, K., Wong, A., and Clausi, D. A. (2018).

Hyperstacknet: A hyper stacked hourglass deep

convolutional neural network architecture for joint

player and stick pose estimation in hockey. In 2018 15th

Conference on Computer and Robot Vision (CRV), pp.

313-320. IEEE.

NHK (Japan Broadcasting Corporation). (2023). Movies of

Grand Sumo, NHK Sports, Available at:

https://www3.nhk.or.jp/sports/special/sumomovies

(Accessed: 10 October 10, 2023)

Rangasamy, K., As’ari, M. A., Rahmad, N. A., and Ghazali,

N. F. (2020). Hockey activity recognition using pre-

trained deep learning model. ICT Express, vol.6, no.3,

pp.170-174.

Shi, L., Zhang, Y., Cheng, J., and Lu, H. (2019). Two-

stream adaptive graph convolutional networks for

skeleton-based action recognition. In Proceedings of

the IEEE/CVF conference on computer vision and

pattern recognition, pp.12026-12035.

Vainstein, J., Manera, J. F., Negri, P., Delrieux, C., and

Maguitman, A. (2014). Modeling video activity with

dynamic phrases and its application to action

recognition in tennis videos. In Progress in Pattern

Recognition, Image Analysis, Computer Vision, and

Applications: 19th Iberoamerican Congress, CIARP

2014, pp.909-916, Springer.

Vats, K., Neher, H., Clausi, D. A., and Zelek, J. (2019).

Two-stream action recognition in ice hockey using

player pose sequences and optical flows. In 2019 16th

Conference on computer and robot vision (CRV),

pp.181-188. IEEE.

Zakharchenko, I. (2020). Basketball Pose-Based Action

Recognition. Ukrainian Catholic University, Faculty of

Applied Sciences: L’viv.

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

408